1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
|
#pragma once
#include <ATen/ATen.h>
#include <ATen/core/ivalue.h>
#include <ATen/core/jit_type.h>
#include <ATen/core/stack.h>
#include <torch/csrc/Export.h>
#include <torch/csrc/jit/ir/ir.h>
#include <list>
#include <map>
#include <unordered_map>
#include <vector>
// We would like to assign each position/axis of a tensor an abstract size
// * For each `tensor` we have a profiled `Value` of a `TensorType` describing
// the properties of the `tensor`.
// * `TensorType` has a property called `symbolic_sizes_` to describe observed
// `tensor.sizes()`
// * `symbolic_sizes_` is a vector of abstract sizes (or
// `std::vector<ShapeSymbol>`) where
// * `ShapeSymbol`at `symbolic_sizes_[i]` describes the size value
// (`Dimension`) at `tensor.sizes()[i]`
// * We may see the same `Dimension` at different positions `i` in
// `tensor.sizes()` or even in different `tensor`
// * First, we would like associate the same `ShapeSymbol` to the same
// `Dimension` across **one** profiling execution or run of a TorchScript
// function.
// * The same `ShapeSymbol`s in different positions of `symbolic_shapes_` in
// possibly different `TensorType`s (i.e. `TensorType`s for different
// profiled values) form an implicit set. The elements of such a set are
// called *dimension locations*.
// * These sets allow us to track how the shapes of input arguments of some
// operation relate to operation's output shapes as the input and output
// shapes might share the same `ShapeSymbol`s
// * For **every** profiling run, we would like to maintain the invariant that
// *the same `ShapeSymbol` is always associated with the same `Dimension`*.
// * To maintain this invariant we merge the profiling information from all
// profiling runs,
// * For every two runs, we iterate over all `symbic_shapes_` and compare
// their `ShapeSymbol`s in the same position.
// * if we observe that for every dimension location that has
// the`ShapeSymbol S1` in run #1 there is **only one** `ShapeSymbol S2` in
// the same dimension location in run #2, we conclude that the invariant
// holds.
// * However, if we observe some dimension locations in run #2 have
// `ShapeSymbol S2` and the other ones have `ShapeSymbol S3` we would like
// to partition the virtual set of dimension locations associated with
// `ShapeSymbol S1` into two new subsets, so the invariant holds.
// * The partitioning works by assigning a new symbol to the dimension
// locations (associated with `ShapeSymbol S1`) that have `ShapeSymbol S2`
// and another new symbol to the dimension locations that have `ShapeSymbol
// S3`. In other words,
// * Subset #1 will consist of the dimension locations that in run #2 have
// `ShapeSymbol S2` and will have `ShapeSymbol S4` in those dimension
// locations
// * Subset #2 will consist of the dimension locations that in run #2 have
// `ShapeSymbol S4` and will have `ShapeSymbol S5` in those dimension
// locations
// * The effective result of merging the profiling information from two runs
// is new `TensorTypes` whose `symbolic_sizes_` /dimension locations have
// either `ShapeSymbol S4` or `ShapeSymbol S5`.
// * Partitioning can be done even before we have seen all the dimension
// locations associated with `ShapeSymbol S1`
// * We use `getSymbolInSet` of `ShapeSymbolTable` to remember all
// `ShapeSymbols` from run #2 we observed in the dimension locations
// associated with `ShapeSymbol S1` .
// * For every `ShapeSymbol` from run #2 in the dimension location
// associated with `ShapeSymbol S1` `getSymbolInSet` returns a symbol
// that we assign to the dimension location in a new TensorType.
// * It's important to point out that the same `ShapeSymbol S2` from run
// #2 in two dimension locations that have different `ShapeSymbol`s in
// run #1 are different! These dimension locations will belong to
// different subsets and have different `ShapeSymbol`s after merge.
// * On the other hand, for the same `ShapeSymbol S2` in two dimension
// locations that have `ShapeSymbol S1` in run #1`getSymbolInSet` will
// return the same symbol.
namespace torch {
namespace jit {
using ::c10::TensorTypePtr;
using Dimension = int64_t;
TORCH_API void RegisterProfilingNode(const std::function<bool(const Node*)>&);
struct ProfilingRecord;
// `SetPartitioningHelper` is used to maintain the following invariant:
// For **every** profiling run, *the same `ShapeSymbol` is always associated
// with the same `Dimension`*.
// while merging the profiling information from multiple runs.
struct SetPartitioningHelper {
std::map<c10::ShapeSymbol, std::map<Dimension, c10::ShapeSymbol>>
sets2subsets_;
// `partitionSetByDimension` partitions a virtual set
// of dimension locations associated with ShapeSymbol `symbol` into subsets.
// Partitioning is equivalent to giving (or renaming) a particular
// dimension location a new `ShapeSymbol`.
// The same `Dimension` value in different dimension locations
// that used to have `symbol` will receive the same
// new `ShapeSymbol`, effectively forming a new set.
c10::ShapeSymbol partitionSetByDimension(
Dimension new_size,
c10::ShapeSymbol symbol) {
auto& dims2symbols = getSetForSymbol(symbol);
if (dims2symbols.count(new_size) == 0) {
auto new_sym = c10::ShapeSymbol::newSymbol();
dims2symbols[new_size] = new_sym;
return new_sym;
}
return dims2symbols[new_size];
}
private:
std::map<Dimension, c10::ShapeSymbol>& getSetForSymbol(c10::ShapeSymbol s) {
auto& set = sets2subsets_[s];
// N.B. adding a mapping { s.static_size(), s }
// makes sure we preserve the fact that
// some dimension values remain the same
// across all profiled runs
if (s.is_static()) {
set.insert({s.static_size(), s});
}
return set;
}
};
// ShapeSymbolTable is used by Interpreter
// to assign dimension values to ShapeSymbols
// and fail a guard if the same symbol
// is assigned more than one dimension value.
struct ShapeSymbolTable {
// N.B. we treat static symbols as always assigned
// to themselves
bool isBound(c10::ShapeSymbol s) {
if (s.is_static()) {
return true;
}
return data_.count(s) != 0;
}
// N.B. we treat static symbols as always assigned
// to themselves
Dimension getValue(c10::ShapeSymbol s) {
if (s.is_static()) {
return s.static_size();
}
return data_[s];
}
void assign(c10::ShapeSymbol s, Dimension v) {
TORCH_INTERNAL_ASSERT(!s.is_static());
data_[s] = v;
}
std::map<c10::ShapeSymbol, Dimension> data_;
// Tries to assign dimension values from `new_sizes` to
// `ShapeSymbol`s `sym_shapes`.
// Returns `true` if every dimension value from `new_sizes`
// can be assigned to the corresponding `ShapeSymbol` from
// `sym_shapes`
// A dimension value can be assigned to a `ShapeSymbol`
// * if the symbol isn't assigned yet any dimension value
// * if the symbol is assigned and its value is equal to
// the dimension value from `new_sizes`
bool bindSymbolicShapes(
at::IntArrayRef new_sizes,
const c10::SymbolicShape& sym_shapes);
};
struct ProfilingRecord {
// N.B. ProfilingRecord's copy and move c-tor are disabled, so we won't
// end up accidentally copying or moving ProfilingRecords whose addresses
// are captured in callbacks_
ProfilingRecord(const ProfilingRecord&) = delete;
ProfilingRecord(ProfilingRecord&&) noexcept = delete;
TORCH_API static std::unique_ptr<ProfilingRecord> instrumentGraph(
const std::shared_ptr<Graph>& graph);
TORCH_API static void removeProfilingNodes(Block* b);
TORCH_API static void removeProfileCounter(Block* b);
std::shared_ptr<Graph> profiled_graph_;
mutable std::mutex mutex_;
size_t profiling_count_;
bool ready() const;
std::shared_ptr<Graph> graph() const {
return profiled_graph_;
}
TORCH_API ProfileIValueOp* createProfileIValueNode(Value* in_val);
TORCH_API ProfileIValueOp* createProfileIValueNode(ArrayRef<Value*> inputs);
private:
ProfileOp* createProfileNode(
const std::function<void(Stack&)>& fp,
at::ArrayRef<Value*> inputs);
void instrumentBlock(Block* block);
void insertShapeProfile(Node* n, size_t offset, const TypePtr& input_type);
ProfilingRecord(std::shared_ptr<Graph> g);
};
} // namespace jit
} // namespace torch
|
